Learning Weather Navigation Skills from Human Pilots Demonstrations using Airborne Radar Data

Weather navigation is one of the essential capabilities for building an autonomous flight navigation system. It is a challenging task for human pilots that requires a great deal of training, background knowledge, and experience. Manually designing such knowledge into an autonomy system, if at all possible, would require large engineering effort and painstaking manual tuning to achieve human-level performance. In this work, we frame weather navigation as an imitation learning problem where human pilot demonstrations are used to learn how to navigate a weather region. We adopt an inverse optimal control formulation which seeks a cost function, i.e., a mapping from weather features to a scalar value cost, under which human pilot demonstrations are optimal. For a given weather navigation problem, the learned cost function is used to generate a costmap that in turn is used by a deterministic planner for navigating the weather region. As a proof of concept we apply this framework to a simplified problem of in-flight weather navigation using airborne radar data. Through a number of training and validation tests we demonstrate its effectiveness in learning from human pilot demonstrations and producing results that are comparable, and in some cases on par, with that of human pilots in generating routes to navigate a variety of weather scenarios. Although the simplified proof of concept presented here focuses on using airborne radar data only, the underlying approach is flexible, and we discuss the remaining challenges and ways of extending it to include other sources of weather data and achieve improved performance.